Medical imaging.
Research in progress.

Research notebook

Progress, with dates.

1. 2 October 2026

   Proposals and source review

   Completed three internally reviewed proposals, candidate-source reviews and a fixed synthetic scaffold. Unknown permissions remain holds on real-data work.

   Planning and engineering checkpoint; no real-data study started.

2. 3 October 2026

   Reproducible synthetic comparisons

   Completed the three toy comparison packages and negative-case evaluation. Recorded unsuccessful outcomes, deterministic repeats and checkpoint reload checks; internal engineering review completed.

   Historical verification: 373 software tests and 10 fixed pilots passed on 3 October.

3. 15 October 2026

   Offline failure analysis

   Audited frozen synthetic results using training and development statistics. Identified signal ceilings, empty-case learning failures and a dose bias contribution of 83.4% of training squared error. This audit itself did not retrain neural models.

   Audit verification: 139 software tests and five scaffold pilots passed. This narrower check did not rerun the full 3 October suite.

4. 15 October 2026 · Follow-up

   One controlled dose experiment

   Completed a frozen two-arm synthetic experiment changing only training-fitted output initialization. Bias and whole-grid development error improved, while exact target-region collapse and worse target D2/V10 endpoints remained. No old test cases were used for this comparison; internal engineering review completed.

   Checkpoint verification: 444 software tests and 10 fixed regression pilots passed. Development results are not fresh held-out or clinical evidence.

5. 18 October 2026

   Architecture and workflow plan

   Prepared first internally reviewed drafts for all three studies, covering source permissions, cohort QA, patient partitions, training-only fitting, controlled comparisons and locked evaluation. The designs remain provisional; clinical source access and expert review are unresolved.

   Documentation-only checkpoint. No new experiments, model fitting or data acquisition; historical synthetic results remain unchanged.

6. 18 October 2026 · Study design

   Study design and data contracts

   Prepared internally reviewed draft study designs, typed row-free dictionaries, exclusion rules and decision checklists for all three studies. A shared contract specifies identifiers, units, missingness and permitted field uses. These describe proposed rules, not populated clinical cohorts; clinical and source decisions remain pending.

   Documentation-only checkpoint. No medical-data downloads, inspection, training or new research experiments were performed.

7. 22 October 2026

   Study scope and source feasibility

   Defined the initial cirrhosis lesion-work-up population for Prism and the original-report, routine adult contrast-CT workflow for Trace. Lumen will select a coherent population through suitable authorized planning data. Source reviews and pre-access specifications document the remaining gaps; no qualifying source or clinical cohort is approved.

   Planning checkpoint, with internal review. Clinical criteria, source rights, reference adequacy and appropriate expert review remain unresolved. No clinical cohort was acquired for analysis and no model was fitted in this planning checkpoint.

8. 23 October 2026

   Offline metadata infrastructure checks

   Completed a bounded radiotherapy metadata-software increment using invented inputs only. All 32 synthetic cases passed in three fresh processes with matching reports, alongside 116 non-fitting regression checks. Historical model findings are unchanged.

   Software verification only: no real source request, medical-data download, clinical inspection or model training in this increment. These checks establish neither source compatibility nor permission, clinical performance or treatment readiness.

9. 25 October 2026

   Source feasibility review

   A bounded public-metadata search across the three studies accepted no new source, cohort, reference or reuse-rights evidence. Some responses were excluded under the review's content limits and other searches returned no results. These outcomes do not establish that suitable sources do not exist or that permission was denied.

   No medical images or clinical files were downloaded and no model was fitted. Source suitability, permissions and clinical review remain unresolved.

10. 26 October 2026

    Synthetic review facility qualification

    Synthetic metadata receipt and review-binding checks passed using invented inputs and simulated review. A separate review-interface qualification stopped at an aggregate layout boundary on its first case, so the facility remains unqualified.

    A possible coordinate correction was documented in a static review; the failure's cause is unknown and the correction has not been implemented or tested. No source or clinical permission, model performance or treatment readiness follows.